Digital Twin Simulation

Multi-echelon Scenario Runner

Autonomy includes a multi-echelon supply chain scenario runner (Retailer, Wholesaler, Distributor, Factory), supporting 2-8 participants per scenario in real time via WebSocket. The scenario runner is the validation substrate behind every agent that lands in production: nothing acts autonomously until it has cleared the twin.

Scenarios surface the dynamics that make supply chains hard to plan: the bullwhip effect (small downstream demand changes amplifying into large upstream order swings), capacity-constrained allocation trade-offs across customer segments, supplier disruption cascades, and the cost of decision latency. These are the regimes where rational individual decisions produce irrational system behaviour, and where the four pillars (agents, conformal, twin, causal) earn their composition.

Mixed Human-AI Scenarios

The real power is in mixed scenarios. Place your team alongside AI agents to see the difference in real-time:

• Human vs Human, Multi-echelon scenario for bullwhip and systems-dynamics education
• Human vs AI, Your team competes against AI agents to see autonomous performance
• Human + AI, Team up with AI agents in mixed roles to build collaboration skills
• AI vs AI, Benchmark different agent strategies against each other

Four Use Cases

1. Employee Training

Scenario-based learning achieves 3-5x higher engagement than traditional supply chain training. New hires experience the bullwhip effect firsthand, understand why autonomous planning matters, and learn to work with AI agents before touching production systems.

2. Agent Validation

Before deploying agents to production, run them through simulation scenarios that stress-test edge cases: demand spikes, supplier disruptions, capacity constraints, quality holds. Verify agent behavior in a risk-free environment with full observability.

3. Confidence Building

Human vs AI competitions demonstrate AI effectiveness in a way that dashboards and reports cannot. When planners see agents consistently outperform them on cost and service level in the simulation, trust in autonomous decisions follows.

4. Continuous Improvement

Human decisions in simulation generate training data for AI agents. Override patterns reveal where human judgment adds value, feeding back into the agent training pipeline for continuous improvement.

"The companies that will lead in the next decade are those that can simulate, test, and validate supply chain decisions before execution. The cost of experimentation in the real world is simply too high."

Deployment Pipeline + Continuous Relearning

The simulation substrate is wired into a 7-step deployment pipeline that brings a new tenant from a seed configuration to live operational plans, plus a continuous CDC relearning loop that keeps the agent stack calibrated against production outcomes.

The 7-step deployment pipeline:

1. Seed configuration, Capacities, BOMs, lanes, calendars, and tenant master data are loaded
2. Deterministic simulation, A baseline trajectory is established without stochastic perturbation
3. Stochastic Monte Carlo, Sampled disruptions stress-test the substrate across the curriculum
4. Training-data conversion, Simulation outputs become training-ready trajectories for the agent stack
5. Model training, Strategic S&OP GraphSAGE, tactical tGNNs, and per-decision TRMs train to the BSC reward
6. Day-1 plans, Initial operational CSVs are generated from the live state
7. Day-2 plans, Continuous-evolution plans are generated as the substrate begins to act

Continuous CDC relearning runs against production outcomes: hourly TRM outcome collection feeds the override-effectiveness posterior; daily causal matching attributes outcomes to agent decisions and human overrides under the Pillar 4 substrate; every-two-hours an Escalation Arbiter evaluates guardrail breaches; weekly data-drift monitoring catches distribution shifts that warrant retraining. Operator overrides in copilot mode are a first-class feedback channel, not a workaround.

"The future of supply chain training isn't classrooms, it's simulation. When planners can experience disruptions in a safe environment, they make better decisions under pressure in the real world."